Modelproof
Always-current AI model facts, and a setup that installs into the tools you already use.
Everyone has access to more models than they can keep track of. Modelproof keeps the facts
straight and puts them where you work:
- The facts — live prices, OpenRouter usage, sourced scores, effort-cost curves, a release
timeline, and what each lab says about its own models, every one with its link and date.
Nothing is ranked, and no model is named for you.
- The instruction package — a short setup for Claude Code, Codex, Cursor, or any tool that
reads
AGENTS.md. It sets which model each helper agent runs on — your choice first, then the
tool's or lab's own published default for that job, otherwise the same model as your main one —
plus a few lines on using them well. Every package includes a checker helper that re-checks
finished work in a fresh context with real evidence and never edits files. Two opt-ins stay off
unless you ask: big-job mode (your main model plans and reviews, the helpers do every piece) and
an example setup someone shared, labeled as one person's own choice. It adds files next to yours
instead of overwriting them, shows you the full plan before it writes anything, and undoes byte
for byte.
Live at lucascashwell3-ai.github.io/modelproof.
Get it
- Paste a prompt — the prompt in
assets/install-prompt.txt (also
on the home page) goes into Claude Code, Codex or Cursor. It reads your setup, asks a few quick
questions, shows every file before it writes, installs on your yes, and prints one undo line.
- The skill —
skills/modelproof-advisor/ runs the same flow. The
prompt runs it once without installing it; to keep it, copy the folder into ~/.claude/skills/
(or your tool's skills folder).
- The board —
board.html builds the same package from a board you lay out, under
Install package.
- MCP server —
mcp/ serves the same facts (prices, releases, sourced quotes) over the
Model Context Protocol.
What the site shows
- Build (the home page) — three taps (your tool, your plan, how you work) draw the setup as a
board of jobs and the model behind each one. Click a box to change its model, then Install copies
the prompt.
- Models — what turning up a model's reasoning-effort dial buys (published cost-accuracy
curves, only for ladders with a named publisher, harness and method), a side-by-side compare of
up to three models, and a track of dated releases by lab. Every section carries the date of its
data and says so when that data is old.
- Teams — what comes after launch.
- Full table (
table.html) — every tracked model with its sources and sourced
quotes, filterable by lab.
- Board (
board.html) — lay out the teams and routines in a company, or your
own set of models, and see the monthly bill and the cost per person and per year.
- How we source facts (
how-we-pick.html) — where every number and quote
comes from.
The data
data/models.json is the single source: models across the major labs and
open-weight vendors, refreshed twice a week by an automated Collect → Judge → Verify pipeline
(GitHub Actions + a research pass with citations; see
automation/jobs/auto-refresh/). Anything the pipeline can't
source cleanly is held in a review issue instead of published. Each model also carries an
availability object — where it can actually be reached (direct API, OpenRouter, AWS Bedrock,
open weights) — derived automatically every Collect run by
scripts/derive-availability.mjs; a field is true only when
sourced, null otherwise, never a guessed negative. data/plans.json holds
seat/subscription pricing for the major chat apps and coding tools (Claude, ChatGPT, Gemini,
Grok, Cursor, GitHub Copilot), refreshed manually (see
scripts/refresh-plans.md) straight from each vendor's own pricing
page — a null price means the page didn't show that number in static HTML, never an invented one.
Honesty rules:
- Pricing traces to official vendor pages (standard tier, USD per 1M tokens).
Verify anything cost-critical against the vendor's own pricing page before relying on it.
- Benchmarks are directional and cited — never treated as truth.
- No ranking. The site never ranks models or names one for a job. Lab and tool quotes
(each with a source link, a verbatim quote and the date checked) are shown as what they are:
that lab's own words about its own models. Every quote is checked against its live source
before publish (
scripts/check-sources.mjs) — a claim that can't
be found on the page it cites is rejected, not softened.
- A blank (—) means "not reliably sourced," never a guess. The validator
(
scripts/validate-data.mjs) blocks any publish that breaks
schema or sourcing rules.
- Independent — not affiliated with, sponsored by, or advertising for any model vendor.
Retired pieces (the old ranking code and its answer key) are kept for history in
archive/.
Run it locally
Pure static site — no build step, no dependencies.
python3 -m http.server 8475 --directory .
# → http://localhost:8475
Tech
Vanilla HTML + CSS + JS, one models.json. Type: Fraunces (wordmark) + Cabinet Grotesk /
Switzer / JetBrains Mono. Restrained, GPU-friendly motion; no build tooling anywhere.
Roadmap
Independent project · not affiliated with any model vendor · built by Lucas Cashwell.
Sibling project: DATproof.
